Why AI coding agents hit 8 shell differences

On November 1, 2026, the AI section of DEV Community published a piece titled "Agent's shell is not your shell: 8 differences that make commands fail". The article argues that when AI coding agents run commands they typically do so inside a non-interactive, non-login shell with output piped rather than attached to a terminal, so startup files and tool behavior end up differing from what a user sees when typing manually.

The author opens with scenarios that readers are likely to recognize. A developer ran nvm use 22 ten minutes earlier and the prompt still shows the expected Node version, yet when the agent is asked to run the tests it reports a syntax error that only an older Node would throw. Or ll comes back as "not found". Or npm test sits there until it times out. Nothing is actually broken, the article stresses; the mismatch comes from the shell environment in which the agent executes commands. The piece then enumerates eight main differences and pairs each one with a command readers can run locally to compare the two environments themselves.

The article is part of the author's Pocket Guides for Developers series, eight short books covering agents, RAG, prompts, observability, events, databases and system design, and it points to the author's separate Hermes IDE project aimed at developers who ship with Claude Code and other AI coding tools. As of the available reporting there is only this single piece, with no follow-up coverage noted yet.